Skip to contentSkip to conversation

perfcounters: IRQ and NMI support on AMD CPUs

PATCH In review Peter Zijlstra 6 participants 16 messages Mar 5, 2009 – Mar 12, 2009 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perfcounters: IRQ and NMI support on AMD CPUs

Subject: perfcounters: IRQ and NMI support on AMD CPUs
From: Peter Zijlstra <[email protected]>
Date: Thu Mar 05 18:08:27 CET 2009

The below completes the K7+ performance counter support.

---
 arch/x86/kernel/cpu/perf_counter.c |  272 +++++++++++++++++++++++++++++++------
 1 file changed, 228 insertions(+), 44 deletions(-)
arch/x86/kernel/cpu/perf_counter.c +228 -44
Index: linux-2.6/arch/x86/kernel/cpu/perf_counter.c
===================================================================
@@ -28,6 +28,7 @@ static bool perf_counters_initialized __
 static int nr_counters_generic __read_mostly;
 static u64 perf_counter_mask __read_mostly;
 static u64 counter_value_mask __read_mostly;
+static int counter_value_bits __read_mostly;
 
 static int nr_counters_fixed __read_mostly;
 
@@ -35,7 +36,9 @@ struct cpu_hw_counters {
 	struct perf_counter	*counters[X86_PMC_IDX_MAX];
 	unsigned long		used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 	unsigned long		interrupts;
-	u64			global_enable;
+	u64			throttle_ctrl;
+	u64			active_mask;
+	int			enabled;
 };
 
 /*
@@ -43,21 +46,28 @@ struct cpu_hw_counters {
  */
 struct pmc_x86_ops {
 	u64		(*save_disable_all)(void);
-	void		(*restore_all)(u64 ctrl);
+	void		(*restore_all)(u64);
+	u64		(*get_status)(u64);
+	void		(*ack_status)(u64);
+	void		(*enable)(int, u64);
+	void		(*disable)(int, u64);
 	unsigned	eventsel;
 	unsigned	perfctr;
-	int		(*event_map)(int event);
+	u64		(*event_map)(int);
+	u64		(*raw_event)(u64);
 	int		max_events;
 };
 
 static struct pmc_x86_ops *pmc_ops;
 
-static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters);
+static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = {
+	.enabled = 1,
+};
 
 /*
  * Intel PerfMon v3. Used on Core2 and later.
  */
-static const int intel_perfmon_event_map[] =
+static const u64 intel_perfmon_event_map[] =
 {
   [PERF_COUNT_CPU_CYCLES]		= 0x003c,
   [PERF_COUNT_INSTRUCTIONS]		= 0x00c0,
@@ -68,15 +78,29 @@ static const int intel_perfmon_event_map
   [PERF_COUNT_BUS_CYCLES]		= 0x013c,
 };
 
-static int pmc_intel_event_map(int event)
+static u64 pmc_intel_event_map(int event)
 {
 	return intel_perfmon_event_map[event];
 }
 
+static u64 pmc_intel_raw_event(u64 event)
+{
+#define CORE_EVNTSEL_EVENT_MASK		0x000000FF
+#define CORE_EVNTSEL_UNIT_MASK		0x0000FF00
+#define CORE_EVNTSEL_COUNTER_MASK	0xFF000000
+
+#define CORE_EVNTSEL_MASK 		\
+	(CORE_EVNTSEL_EVENT_MASK |	\
+	 CORE_EVNTSEL_UNIT_MASK  |	\
+	 CORE_EVNTSEL_COUNTER_MASK)
+
+	return event & CORE_EVNTSEL_MASK;
+}
+
 /*
  * AMD Performance Monitor K7 and later.
  */
-static const int amd_perfmon_event_map[] =
+static const u64 amd_perfmon_event_map[] =
 {
   [PERF_COUNT_CPU_CYCLES]		= 0x0076,
   [PERF_COUNT_INSTRUCTIONS]		= 0x00c0,
@@ -86,11 +110,25 @@ static const int amd_perfmon_event_map[]
   [PERF_COUNT_BRANCH_MISSES]		= 0x00c5,
 };
 
-static int pmc_amd_event_map(int event)
+static u64 pmc_amd_event_map(int event)
 {
 	return amd_perfmon_event_map[event];
 }
 
+static u64 pmc_amd_raw_event(u64 event)
+{
+#define K7_EVNTSEL_EVENT_MASK	0x7000000FF
+#define K7_EVNTSEL_UNIT_MASK	0x00000FF00
+#define K7_EVNTSEL_COUNTER_MASK	0x0FF000000
+
+#define K7_EVNTSEL_MASK			\
+	(K7_EVNTSEL_EVENT_MASK |	\
+	 K7_EVNTSEL_UNIT_MASK  |	\
+	 K7_EVNTSEL_COUNTER_MASK)
+
+	return event & K7_EVNTSEL_MASK;
+}
+
 /*
  * Propagate counter elapsed time into the generic counter.
  * Can only be executed on the CPU where the counter is active.
@@ -179,7 +217,7 @@ static int __hw_perf_counter_init(struct
 	 * Raw event type provide the config in the event structure
 	 */
 	if (hw_event->raw) {
-		hwc->config |= hw_event->type;
+		hwc->config |= pmc_ops->raw_event(hw_event->type);
 	} else {
 		if (hw_event->type >= pmc_ops->max_events)
 			return -EINVAL;
@@ -205,18 +243,24 @@ static u64 pmc_intel_save_disable_all(vo
 
 static u64 pmc_amd_save_disable_all(void)
 {
-	int idx;
-	u64 val, ctrl = 0;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	int enabled, idx;
+
+	enabled = cpuc->enabled;
+	cpuc->enabled = 0;
+	barrier();
 
 	for (idx = 0; idx < nr_counters_generic; idx++) {
+		u64 val;
+
 		rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
-		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
-			ctrl |= (1 << idx);
-		val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
-		wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE) {
+			val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
+			wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		}
 	}
 
-	return ctrl;
+	return enabled;
 }
 
 u64 hw_perf_save_disable(void)
@@ -226,6 +270,9 @@ u64 hw_perf_save_disable(void)
 
 	return pmc_ops->save_disable_all();
 }
+/*
+ * Exported because of ACPI idle
+ */
 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
 
 static void pmc_intel_restore_all(u64 ctrl)
@@ -235,11 +282,18 @@ static void pmc_intel_restore_all(u64 ct
 
 static void pmc_amd_restore_all(u64 ctrl)
 {
-	u64 val;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 	int idx;
 
+	cpuc->enabled = ctrl;
+	barrier();
+	if (!ctrl)
+		return;
+
 	for (idx = 0; idx < nr_counters_generic; idx++) {
-		if (ctrl & (1 << idx)) {
+		if (test_bit(idx, (unsigned long *)&cpuc->active_mask)) {
+			u64 val;
+
 			rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
 			val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 			wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
@@ -254,8 +308,112 @@ void hw_perf_restore(u64 ctrl)
 
 	pmc_ops->restore_all(ctrl);
 }
+/*
+ * Exported because of ACPI idle
+ */
 EXPORT_SYMBOL_GPL(hw_perf_restore);
 
+static u64 pmc_intel_get_status(u64 mask)
+{
+	u64 status;
+
+	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+
+	return status;
+}
+
+static u64 pmc_amd_get_status(u64 mask)
+{
+	u64 status = 0;
+	int idx;
+
+	for (idx = 0; idx < nr_counters_generic; idx++) {
+		s64 val;
+
+		if (!(mask & (1 << idx)))
+			continue;
+
+		rdmsrl(MSR_K7_PERFCTR0 + idx, val);
+		val <<= (64 - counter_value_bits);
+		if (val >= 0)
+			status |= (1 << idx);
+	}
+
+	return status;
+}
+
+static u64 hw_perf_get_status(u64 mask)
+{
+	if (unlikely(!perf_counters_initialized))
+		return 0;
+
+	return pmc_ops->get_status(mask);
+}
+
+static void pmc_intel_ack_status(u64 ack)
+{
+	wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
+}
+
+static void pmc_amd_ack_status(u64 ack)
+{
+}
+
+static void hw_perf_ack_status(u64 ack)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->ack_status(ack);
+}
+
+static void pmc_intel_enable(int idx, u64 config)
+{
+	wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx,
+			config | ARCH_PERFMON_EVENTSEL0_ENABLE);
+}
+
+static void pmc_amd_enable(int idx, u64 config)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+
+	set_bit(idx, (unsigned long *)&cpuc->active_mask);
+	if (cpuc->enabled)
+		config |= ARCH_PERFMON_EVENTSEL0_ENABLE;
+
+	wrmsrl(MSR_K7_EVNTSEL0 + idx, config);
+}
+
+static void hw_perf_enable(int idx, u64 config)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->enable(idx, config);
+}
+
+static void pmc_intel_disable(int idx, u64 config)
+{
+	wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, config);
+}
+
+static void pmc_amd_disable(int idx, u64 config)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+
+	clear_bit(idx, (unsigned long *)&cpuc->active_mask);
+	wrmsrl(MSR_K7_EVNTSEL0 + idx, config);
+
+}
+
+static void hw_perf_disable(int idx, u64 config)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->disable(idx, config);
+}
+
 static inline void
 __pmc_fixed_disable(struct perf_counter *counter,
 		    struct hw_perf_counter *hwc, unsigned int __idx)
@@ -278,7 +436,7 @@ __pmc_generic_disable(struct perf_counte
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
 		__pmc_fixed_disable(counter, hwc, idx);
 	else
-		wrmsr_safe(hwc->config_base + idx, hwc->config, 0);
+		hw_perf_disable(idx, hwc->config);
 }
 
 static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
@@ -354,8 +512,7 @@ __pmc_generic_enable(struct perf_counter
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
 		__pmc_fixed_enable(counter, hwc, idx);
 	else
-		wrmsr(hwc->config_base + idx,
-		      hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE, 0);
+		hw_perf_enable(idx, hwc->config);
 }
 
 static int
@@ -567,22 +724,20 @@ perf_handle_group(struct perf_counter *s
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
  */
-static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
+static int __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
 {
 	int bit, cpu = smp_processor_id();
 	u64 ack, status;
 	struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
+	int ret = 0;
 
-	rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+	cpuc->throttle_ctrl = hw_perf_save_disable();
 
-	/* Disable counters globally */
-	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
-	ack_APIC_irq();
-
-	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+	status = hw_perf_get_status(cpuc->throttle_ctrl);
 	if (!status)
 		goto out;
 
+	ret = 1;
 again:
 	inc_irq_stat(apic_perf_irqs);
 	ack = status;
@@ -618,12 +773,12 @@ again:
 		}
 	}
 
-	wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
+	hw_perf_ack_status(ack);
 
 	/*
 	 * Repeat if there is more work to be done:
 	 */
-	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+	status = hw_perf_get_status(cpuc->throttle_ctrl);
 	if (status)
 		goto again;
 out:
@@ -631,32 +786,27 @@ out:
 	 * Restore - do not reenable when global enable is off or throttled:
 	 */
 	if (++cpuc->interrupts < PERFMON_MAX_INTERRUPTS)
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+		hw_perf_restore(cpuc->throttle_ctrl);
+
+	return ret;
 }
 
 void perf_counter_unthrottle(void)
 {
 	struct cpu_hw_counters *cpuc;
-	u64 global_enable;
 
 	if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 		return;
 
-	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
-		return;
-
 	if (unlikely(!perf_counters_initialized))
 		return;
 
-	cpuc = &per_cpu(cpu_hw_counters, smp_processor_id());
+	cpuc = &__get_cpu_var(cpu_hw_counters);
 	if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
 		if (printk_ratelimit())
 			printk(KERN_WARNING "PERFMON: max interrupts exceeded!\n");
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+		hw_perf_restore(cpuc->throttle_ctrl);
 	}
-	rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, global_enable);
-	if (unlikely(cpuc->global_enable && !global_enable))
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
 	cpuc->interrupts = 0;
 }
 
@@ -664,8 +814,8 @@ void smp_perf_counter_interrupt(struct p
 {
 	irq_enter();
 	apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
+	ack_APIC_irq();
 	__smp_perf_counter_interrupt(regs, 0);
-
 	irq_exit();
 }
 
@@ -722,16 +872,23 @@ perf_counter_nmi_handler(struct notifier
 {
 	struct die_args *args = __args;
 	struct pt_regs *regs;
+	int ret;
+
+	switch (cmd) {
+	case DIE_NMI:
+	case DIE_NMI_IPI:
+		break;
 
-	if (likely(cmd != DIE_NMI_IPI))
+	default:
 		return NOTIFY_DONE;
+	}
 
 	regs = args->regs;
 
 	apic_write(APIC_LVTPC, APIC_DM_NMI);
-	__smp_perf_counter_interrupt(regs, 1);
+	ret = __smp_perf_counter_interrupt(regs, 1);
 
-	return NOTIFY_STOP;
+	return ret ? NOTIFY_STOP : NOTIFY_OK;
 }
 
 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
@@ -743,18 +900,28 @@ static __read_mostly struct notifier_blo
 static struct pmc_x86_ops pmc_intel_ops = {
 	.save_disable_all	= pmc_intel_save_disable_all,
 	.restore_all		= pmc_intel_restore_all,
+	.get_status		= pmc_intel_get_status,
+	.ack_status		= pmc_intel_ack_status,
+	.enable			= pmc_intel_enable,
+	.disable		= pmc_intel_disable,
 	.eventsel		= MSR_ARCH_PERFMON_EVENTSEL0,
 	.perfctr		= MSR_ARCH_PERFMON_PERFCTR0,
 	.event_map		= pmc_intel_event_map,
+	.raw_event		= pmc_intel_raw_event,
 	.max_events		= ARRAY_SIZE(intel_perfmon_event_map),
 };
 
 static struct pmc_x86_ops pmc_amd_ops = {
 	.save_disable_all	= pmc_amd_save_disable_all,
 	.restore_all		= pmc_amd_restore_all,
+	.get_status		= pmc_amd_get_status,
+	.ack_status		= pmc_amd_ack_status,
+	.enable			= pmc_amd_enable,
+	.disable		= pmc_amd_disable,
 	.eventsel		= MSR_K7_EVNTSEL0,
 	.perfctr		= MSR_K7_PERFCTR0,
 	.event_map		= pmc_amd_event_map,
+	.raw_event		= pmc_amd_raw_event,
 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
 };
 
@@ -787,8 +954,25 @@ static struct pmc_x86_ops *pmc_intel_ini
 
 static struct pmc_x86_ops *pmc_amd_init(void)
 {
+	u64 old;
+	int bits;
+
 	nr_counters_generic = 4;
 	nr_counters_fixed = 0;
+	counter_value_mask = ~0ULL;
+
+	rdmsrl(MSR_K7_PERFCTR0, old);
+	wrmsrl(MSR_K7_PERFCTR0, counter_value_mask);
+	/*
+	 * read the truncated mask
+	 */
+	rdmsrl(MSR_K7_PERFCTR0, counter_value_mask);
+	wrmsrl(MSR_K7_PERFCTR0, old);
+
+	bits = 32 + fls(counter_value_mask >> 32);
+	if (bits == 32)
+		bits = fls((u32)counter_value_mask);
+	counter_value_bits = bits;
 
 	pr_info("AMD Performance Monitoring support detected.\n");
 
#1 lore raw
Quoted 10 lines · Peter Zijlstra

Subject: perfcounters: IRQ and NMI support on AMD CPUs
From: Peter Zijlstra <[email protected]>
Date: Thu Mar 05 18:08:27 CET 2009

The below completes the K7+ performance counter support.

---
 arch/x86/kernel/cpu/perf_counter.c |  272 +++++++++++++++++++++++++++++++------
 1 file changed, 228 insertions(+), 44 deletions(-)

Great work, thanks Peter! I've applied it to
tip:perfcounters/core and tip:master.

Ingo

#2 lore raw
In reply to Peter Zijlstra
Commit-ID:  b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Gitweb:     http://git.kernel.org/tip/b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Author:     "Peter Zijlstra" <[email protected]>
AuthorDate: Thu, 5 Mar 2009 18:08:27 +0100
Commit:     Ingo Molnar <[email protected]>
CommitDate: Thu, 5 Mar 2009 18:25:16 +0100

perfcounters: IRQ and NMI support on AMD CPUs

The below completes the K7+ performance counter support:

- IRQ support
- NMI support

KernelTop output works now as well.

Cc Jaswinder Singh Rajput <[email protected]>
Cc Paul Mackerras <[email protected]>

LKML-Reference: <1236273633.5187.286.camel@laptop>

---
 arch/x86/kernel/cpu/perf_counter.c |  272 ++++++++++++++++++++++++++++++------
 1 files changed, 228 insertions(+), 44 deletions(-)
arch/x86/kernel/cpu/perf_counter.c +228 -44
index 3b65f19..6ebe9ab 100644
@@ -28,6 +28,7 @@ static bool perf_counters_initialized __read_mostly;
 static int nr_counters_generic __read_mostly;
 static u64 perf_counter_mask __read_mostly;
 static u64 counter_value_mask __read_mostly;
+static int counter_value_bits __read_mostly;
 
 static int nr_counters_fixed __read_mostly;
 
@@ -35,7 +36,9 @@ struct cpu_hw_counters {
 	struct perf_counter	*counters[X86_PMC_IDX_MAX];
 	unsigned long		used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 	unsigned long		interrupts;
-	u64			global_enable;
+	u64			throttle_ctrl;
+	u64			active_mask;
+	int			enabled;
 };
 
 /*
@@ -43,21 +46,28 @@ struct cpu_hw_counters {
  */
 struct pmc_x86_ops {
 	u64		(*save_disable_all)(void);
-	void		(*restore_all)(u64 ctrl);
+	void		(*restore_all)(u64);
+	u64		(*get_status)(u64);
+	void		(*ack_status)(u64);
+	void		(*enable)(int, u64);
+	void		(*disable)(int, u64);
 	unsigned	eventsel;
 	unsigned	perfctr;
-	int		(*event_map)(int event);
+	u64		(*event_map)(int);
+	u64		(*raw_event)(u64);
 	int		max_events;
 };
 
 static struct pmc_x86_ops *pmc_ops;
 
-static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters);
+static DEFINE_PER_CPU(struct cpu_hw_counters, cpu_hw_counters) = {
+	.enabled = 1,
+};
 
 /*
  * Intel PerfMon v3. Used on Core2 and later.
  */
-static const int intel_perfmon_event_map[] =
+static const u64 intel_perfmon_event_map[] =
 {
   [PERF_COUNT_CPU_CYCLES]		= 0x003c,
   [PERF_COUNT_INSTRUCTIONS]		= 0x00c0,
@@ -68,15 +78,29 @@ static const int intel_perfmon_event_map[] =
   [PERF_COUNT_BUS_CYCLES]		= 0x013c,
 };
 
-static int pmc_intel_event_map(int event)
+static u64 pmc_intel_event_map(int event)
 {
 	return intel_perfmon_event_map[event];
 }
 
+static u64 pmc_intel_raw_event(u64 event)
+{
+#define CORE_EVNTSEL_EVENT_MASK		0x000000FF
+#define CORE_EVNTSEL_UNIT_MASK		0x0000FF00
+#define CORE_EVNTSEL_COUNTER_MASK	0xFF000000
+
+#define CORE_EVNTSEL_MASK 		\
+	(CORE_EVNTSEL_EVENT_MASK |	\
+	 CORE_EVNTSEL_UNIT_MASK  |	\
+	 CORE_EVNTSEL_COUNTER_MASK)
+
+	return event & CORE_EVNTSEL_MASK;
+}
+
 /*
  * AMD Performance Monitor K7 and later.
  */
-static const int amd_perfmon_event_map[] =
+static const u64 amd_perfmon_event_map[] =
 {
   [PERF_COUNT_CPU_CYCLES]		= 0x0076,
   [PERF_COUNT_INSTRUCTIONS]		= 0x00c0,
@@ -86,11 +110,25 @@ static const int amd_perfmon_event_map[] =
   [PERF_COUNT_BRANCH_MISSES]		= 0x00c5,
 };
 
-static int pmc_amd_event_map(int event)
+static u64 pmc_amd_event_map(int event)
 {
 	return amd_perfmon_event_map[event];
 }
 
+static u64 pmc_amd_raw_event(u64 event)
+{
+#define K7_EVNTSEL_EVENT_MASK	0x7000000FF
+#define K7_EVNTSEL_UNIT_MASK	0x00000FF00
+#define K7_EVNTSEL_COUNTER_MASK	0x0FF000000
+
+#define K7_EVNTSEL_MASK			\
+	(K7_EVNTSEL_EVENT_MASK |	\
+	 K7_EVNTSEL_UNIT_MASK  |	\
+	 K7_EVNTSEL_COUNTER_MASK)
+
+	return event & K7_EVNTSEL_MASK;
+}
+
 /*
  * Propagate counter elapsed time into the generic counter.
  * Can only be executed on the CPU where the counter is active.
@@ -179,7 +217,7 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
 	 * Raw event type provide the config in the event structure
 	 */
 	if (hw_event->raw) {
-		hwc->config |= hw_event->type;
+		hwc->config |= pmc_ops->raw_event(hw_event->type);
 	} else {
 		if (hw_event->type >= pmc_ops->max_events)
 			return -EINVAL;
@@ -205,18 +243,24 @@ static u64 pmc_intel_save_disable_all(void)
 
 static u64 pmc_amd_save_disable_all(void)
 {
-	int idx;
-	u64 val, ctrl = 0;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	int enabled, idx;
+
+	enabled = cpuc->enabled;
+	cpuc->enabled = 0;
+	barrier();
 
 	for (idx = 0; idx < nr_counters_generic; idx++) {
+		u64 val;
+
 		rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
-		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
-			ctrl |= (1 << idx);
-		val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
-		wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE) {
+			val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
+			wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		}
 	}
 
-	return ctrl;
+	return enabled;
 }
 
 u64 hw_perf_save_disable(void)
@@ -226,6 +270,9 @@ u64 hw_perf_save_disable(void)
 
 	return pmc_ops->save_disable_all();
 }
+/*
+ * Exported because of ACPI idle
+ */
 EXPORT_SYMBOL_GPL(hw_perf_save_disable);
 
 static void pmc_intel_restore_all(u64 ctrl)
@@ -235,11 +282,18 @@ static void pmc_intel_restore_all(u64 ctrl)
 
 static void pmc_amd_restore_all(u64 ctrl)
 {
-	u64 val;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
 	int idx;
 
+	cpuc->enabled = ctrl;
+	barrier();
+	if (!ctrl)
+		return;
+
 	for (idx = 0; idx < nr_counters_generic; idx++) {
-		if (ctrl & (1 << idx)) {
+		if (test_bit(idx, (unsigned long *)&cpuc->active_mask)) {
+			u64 val;
+
 			rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
 			val |= ARCH_PERFMON_EVENTSEL0_ENABLE;
 			wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
@@ -254,8 +308,112 @@ void hw_perf_restore(u64 ctrl)
 
 	pmc_ops->restore_all(ctrl);
 }
+/*
+ * Exported because of ACPI idle
+ */
 EXPORT_SYMBOL_GPL(hw_perf_restore);
 
+static u64 pmc_intel_get_status(u64 mask)
+{
+	u64 status;
+
+	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+
+	return status;
+}
+
+static u64 pmc_amd_get_status(u64 mask)
+{
+	u64 status = 0;
+	int idx;
+
+	for (idx = 0; idx < nr_counters_generic; idx++) {
+		s64 val;
+
+		if (!(mask & (1 << idx)))
+			continue;
+
+		rdmsrl(MSR_K7_PERFCTR0 + idx, val);
+		val <<= (64 - counter_value_bits);
+		if (val >= 0)
+			status |= (1 << idx);
+	}
+
+	return status;
+}
+
+static u64 hw_perf_get_status(u64 mask)
+{
+	if (unlikely(!perf_counters_initialized))
+		return 0;
+
+	return pmc_ops->get_status(mask);
+}
+
+static void pmc_intel_ack_status(u64 ack)
+{
+	wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
+}
+
+static void pmc_amd_ack_status(u64 ack)
+{
+}
+
+static void hw_perf_ack_status(u64 ack)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->ack_status(ack);
+}
+
+static void pmc_intel_enable(int idx, u64 config)
+{
+	wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx,
+			config | ARCH_PERFMON_EVENTSEL0_ENABLE);
+}
+
+static void pmc_amd_enable(int idx, u64 config)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+
+	set_bit(idx, (unsigned long *)&cpuc->active_mask);
+	if (cpuc->enabled)
+		config |= ARCH_PERFMON_EVENTSEL0_ENABLE;
+
+	wrmsrl(MSR_K7_EVNTSEL0 + idx, config);
+}
+
+static void hw_perf_enable(int idx, u64 config)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->enable(idx, config);
+}
+
+static void pmc_intel_disable(int idx, u64 config)
+{
+	wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, config);
+}
+
+static void pmc_amd_disable(int idx, u64 config)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+
+	clear_bit(idx, (unsigned long *)&cpuc->active_mask);
+	wrmsrl(MSR_K7_EVNTSEL0 + idx, config);
+
+}
+
+static void hw_perf_disable(int idx, u64 config)
+{
+	if (unlikely(!perf_counters_initialized))
+		return;
+
+	pmc_ops->disable(idx, config);
+}
+
 static inline void
 __pmc_fixed_disable(struct perf_counter *counter,
 		    struct hw_perf_counter *hwc, unsigned int __idx)
@@ -278,7 +436,7 @@ __pmc_generic_disable(struct perf_counter *counter,
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
 		__pmc_fixed_disable(counter, hwc, idx);
 	else
-		wrmsr_safe(hwc->config_base + idx, hwc->config, 0);
+		hw_perf_disable(idx, hwc->config);
 }
 
 static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
@@ -354,8 +512,7 @@ __pmc_generic_enable(struct perf_counter *counter,
 	if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
 		__pmc_fixed_enable(counter, hwc, idx);
 	else
-		wrmsr(hwc->config_base + idx,
-		      hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE, 0);
+		hw_perf_enable(idx, hwc->config);
 }
 
 static int
@@ -567,22 +724,20 @@ perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
  */
-static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
+static int __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
 {
 	int bit, cpu = smp_processor_id();
 	u64 ack, status;
 	struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
+	int ret = 0;
 
-	rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
-
-	/* Disable counters globally */
-	wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
-	ack_APIC_irq();
+	cpuc->throttle_ctrl = hw_perf_save_disable();
 
-	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+	status = hw_perf_get_status(cpuc->throttle_ctrl);
 	if (!status)
 		goto out;
 
+	ret = 1;
 again:
 	inc_irq_stat(apic_perf_irqs);
 	ack = status;
@@ -618,12 +773,12 @@ again:
 		}
 	}
 
-	wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
+	hw_perf_ack_status(ack);
 
 	/*
 	 * Repeat if there is more work to be done:
 	 */
-	rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
+	status = hw_perf_get_status(cpuc->throttle_ctrl);
 	if (status)
 		goto again;
 out:
@@ -631,32 +786,27 @@ out:
 	 * Restore - do not reenable when global enable is off or throttled:
 	 */
 	if (++cpuc->interrupts < PERFMON_MAX_INTERRUPTS)
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+		hw_perf_restore(cpuc->throttle_ctrl);
+
+	return ret;
 }
 
 void perf_counter_unthrottle(void)
 {
 	struct cpu_hw_counters *cpuc;
-	u64 global_enable;
 
 	if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
 		return;
 
-	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
-		return;
-
 	if (unlikely(!perf_counters_initialized))
 		return;
 
-	cpuc = &per_cpu(cpu_hw_counters, smp_processor_id());
+	cpuc = &__get_cpu_var(cpu_hw_counters);
 	if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
 		if (printk_ratelimit())
 			printk(KERN_WARNING "PERFMON: max interrupts exceeded!\n");
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+		hw_perf_restore(cpuc->throttle_ctrl);
 	}
-	rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, global_enable);
-	if (unlikely(cpuc->global_enable && !global_enable))
-		wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
 	cpuc->interrupts = 0;
 }
 
@@ -664,8 +814,8 @@ void smp_perf_counter_interrupt(struct pt_regs *regs)
 {
 	irq_enter();
 	apic_write(APIC_LVTPC, LOCAL_PERF_VECTOR);
+	ack_APIC_irq();
 	__smp_perf_counter_interrupt(regs, 0);
-
 	irq_exit();
 }
 
@@ -722,16 +872,23 @@ perf_counter_nmi_handler(struct notifier_block *self,
 {
 	struct die_args *args = __args;
 	struct pt_regs *regs;
+	int ret;
+
+	switch (cmd) {
+	case DIE_NMI:
+	case DIE_NMI_IPI:
+		break;
 
-	if (likely(cmd != DIE_NMI_IPI))
+	default:
 		return NOTIFY_DONE;
+	}
 
 	regs = args->regs;
 
 	apic_write(APIC_LVTPC, APIC_DM_NMI);
-	__smp_perf_counter_interrupt(regs, 1);
+	ret = __smp_perf_counter_interrupt(regs, 1);
 
-	return NOTIFY_STOP;
+	return ret ? NOTIFY_STOP : NOTIFY_OK;
 }
 
 static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
@@ -743,18 +900,28 @@ static __read_mostly struct notifier_block perf_counter_nmi_notifier = {
 static struct pmc_x86_ops pmc_intel_ops = {
 	.save_disable_all	= pmc_intel_save_disable_all,
 	.restore_all		= pmc_intel_restore_all,
+	.get_status		= pmc_intel_get_status,
+	.ack_status		= pmc_intel_ack_status,
+	.enable			= pmc_intel_enable,
+	.disable		= pmc_intel_disable,
 	.eventsel		= MSR_ARCH_PERFMON_EVENTSEL0,
 	.perfctr		= MSR_ARCH_PERFMON_PERFCTR0,
 	.event_map		= pmc_intel_event_map,
+	.raw_event		= pmc_intel_raw_event,
 	.max_events		= ARRAY_SIZE(intel_perfmon_event_map),
 };
 
 static struct pmc_x86_ops pmc_amd_ops = {
 	.save_disable_all	= pmc_amd_save_disable_all,
 	.restore_all		= pmc_amd_restore_all,
+	.get_status		= pmc_amd_get_status,
+	.ack_status		= pmc_amd_ack_status,
+	.enable			= pmc_amd_enable,
+	.disable		= pmc_amd_disable,
 	.eventsel		= MSR_K7_EVNTSEL0,
 	.perfctr		= MSR_K7_PERFCTR0,
 	.event_map		= pmc_amd_event_map,
+	.raw_event		= pmc_amd_raw_event,
 	.max_events		= ARRAY_SIZE(amd_perfmon_event_map),
 };
 
@@ -787,8 +954,25 @@ static struct pmc_x86_ops *pmc_intel_init(void)
 
 static struct pmc_x86_ops *pmc_amd_init(void)
 {
+	u64 old;
+	int bits;
+
 	nr_counters_generic = 4;
 	nr_counters_fixed = 0;
+	counter_value_mask = ~0ULL;
+
+	rdmsrl(MSR_K7_PERFCTR0, old);
+	wrmsrl(MSR_K7_PERFCTR0, counter_value_mask);
+	/*
+	 * read the truncated mask
+	 */
+	rdmsrl(MSR_K7_PERFCTR0, counter_value_mask);
+	wrmsrl(MSR_K7_PERFCTR0, old);
+
+	bits = 32 + fls(counter_value_mask >> 32);
+	if (bits == 32)
+		bits = fls((u32)counter_value_mask);
+	counter_value_bits = bits;
 
 	pr_info("AMD Performance Monitoring support detected.\n");
 
#5 lore raw
In reply to Peter Zijlstra

Hello Peter,

Ahh, so that day you was getting triple fault in this ?

Quoted 11 lines · Peter Zijlstra

Subject: perfcounters: IRQ and NMI support on AMD CPUs
From: Peter Zijlstra <[email protected]>
Date: Thu Mar 05 18:08:27 CET 2009

The below completes the K7+ performance counter support.

---
 arch/x86/kernel/cpu/perf_counter.c |  272 +++++++++++++++++++++++++++++++------
 1 file changed, 228 insertions(+), 44 deletions(-)

It was in my TODO list, but I am glad that you finished it :-)

Any how with this patch on my 64 bit AMD box I am getting:

[jaswinder@hpdv5 perfcounters]$ ./perfstat -e 0,1,2,3 ls > /dev/null

Performance counter stats for 'ls':

perfstat: perfstat.c:415: main: Assertion `res == sizeof(single_count)' failed.
Aborted
[jaswinder@hpdv5 perfcounters]$

[root@hpdv5 perfcounters]# ./kerneltop
KernelTop refresh period: 2 seconds
kerneltop: kerneltop.c:883: main: Assertion `res == sizeof(ip)' failed. Aborted
[root@hpdv5 perfcounters]#

Thanks,
--
JSR

#6 lore raw
Quoted 30 lines · Jaswinder Singh Rajput

Hello Peter,

Ahh, so that day you was getting triple fault in this ?

Quoted 11 lines · Peter Zijlstra

Subject: perfcounters: IRQ and NMI support on AMD CPUs
From: Peter Zijlstra <[email protected]>
Date: Thu Mar 05 18:08:27 CET 2009

The below completes the K7+ performance counter support.

---
 arch/x86/kernel/cpu/perf_counter.c |  272 +++++++++++++++++++++++++++++++------
 1 file changed, 228 insertions(+), 44 deletions(-)

It was in my TODO list, but I am glad that you finished it :-)

Any how with this patch on my 64 bit AMD box I am getting:

[jaswinder@hpdv5 perfcounters]$ ./perfstat -e 0,1,2,3 ls > /dev/null

Performance counter stats for 'ls':

perfstat: perfstat.c:415: main: Assertion `res == sizeof(single_count)' failed.
Aborted
[jaswinder@hpdv5 perfcounters]$

Quoted 5 lines

[root@hpdv5 perfcounters]# ./kerneltop
KernelTop refresh period: 2 seconds
kerneltop: kerneltop.c:883: main: Assertion `res == sizeof(ip)' failed. Aborted
[root@hpdv5 perfcounters]#

Ah, I think you will need a fresh copy of both utilities, Paul recently changed the ABI.

http://people.redhat.com/mingo/perfcounters/

Contains the latest IIRC

#7 lore raw
Quoted 19 lines · Peter Zijlstra
Quoted 12 lines · Jaswinder Singh Rajput

Any how with this patch on my 64 bit AMD box I am getting:

[jaswinder@hpdv5 perfcounters]$ ./perfstat -e 0,1,2,3 ls > /dev/null

Performance counter stats for 'ls':

perfstat: perfstat.c:415: main: Assertion `res == sizeof(single_count)' failed.
Aborted
[jaswinder@hpdv5 perfcounters]$

Quoted 5 lines

[root@hpdv5 perfcounters]# ./kerneltop
KernelTop refresh period: 2 seconds
kerneltop: kerneltop.c:883: main: Assertion `res == sizeof(ip)' failed. Aborted
[root@hpdv5 perfcounters]#

Quoted 7 lines

Ah, I think you will need a fresh copy of both utilities, Paul recently changed the ABI.

http://people.redhat.com/mingo/perfcounters/

Contains the latest IIRC

With fresh copy, perfstat is working.

But kerneltop is still not working and I am not getting any NMI Interrupts.

[root@hpdv5 new]# ./kerneltop -e 3 -c 100 -e 2

------------------------------------------------------------------------------
KernelTop: 0 irqs/sec kernel:nan% [NMI, cache-misses/cache-refs], (all, 2 CPUs)
------------------------------------------------------------------------------

  weight     events         RIP          kernel function
  ______     ______   ________________   _______________

--
JSR

#8 lore raw
Quoted 23 lines · Jaswinder Singh Rajput
Quoted 22 lines · Peter Zijlstra
Quoted 12 lines · Jaswinder Singh Rajput

Any how with this patch on my 64 bit AMD box I am getting:

[jaswinder@hpdv5 perfcounters]$ ./perfstat -e 0,1,2,3 ls > /dev/null

Performance counter stats for 'ls':

perfstat: perfstat.c:415: main: Assertion `res == sizeof(single_count)' failed.
Aborted
[jaswinder@hpdv5 perfcounters]$

Quoted 5 lines

[root@hpdv5 perfcounters]# ./kerneltop
KernelTop refresh period: 2 seconds
kerneltop: kerneltop.c:883: main: Assertion `res == sizeof(ip)' failed. Aborted
[root@hpdv5 perfcounters]#

Ah, I think you will need a fresh copy of both utilities, Paul recently changed the ABI.

Quoted 8 lines
Quoted 4 lines

http://people.redhat.com/mingo/perfcounters/

Contains the latest IIRC

With fresh copy, perfstat is working.

But kerneltop is still not working and I am not getting any NMI Interrupts.

What does your dmegs look like?

Mine read like:

[    0.012002] AMD Performance Monitoring support detected.
[    0.013002] ... num counters:    4
[    0.014001] ... value mask:      0000ffffffffffff
[    0.015001] ... fixed counters:  0
[    0.016001] ... counter mask:    000000000000000f

and its generating IRQs:

CNT: 168612 166040 Performance counter interrupts

#10 lore raw
In reply to Jaswinder Singh Rajput
Quoted 1 line · Jaswinder Singh Rajput

But kerneltop is still not working and I am not getting any NMI Interrupts.

Does the below work for you?

The BKGD suggests its 48 for all existing machines (it certainly is for mine).

arch/x86/kernel/cpu/perf_counter.c +2 -14
Index: linux-2.6/arch/x86/kernel/cpu/perf_counter.c
===================================================================
@@ -959,20 +959,8 @@ static struct pmc_x86_ops *pmc_amd_init(
 
 	nr_counters_generic = 4;
 	nr_counters_fixed = 0;
-	counter_value_mask = ~0ULL;
-
-	rdmsrl(MSR_K7_PERFCTR0, old);
-	wrmsrl(MSR_K7_PERFCTR0, counter_value_mask);
-	/*
-	 * read the truncated mask
-	 */
-	rdmsrl(MSR_K7_PERFCTR0, counter_value_mask);
-	wrmsrl(MSR_K7_PERFCTR0, old);
-
-	bits = 32 + fls(counter_value_mask >> 32);
-	if (bits == 32)
-		bits = fls((u32)counter_value_mask);
-	counter_value_bits = bits;
+	counter_value_mask = 0x0000FFFFFFFFFFFFULL;
+	counter_value_bits = 48;
 
 	pr_info("AMD Performance Monitoring support detected.\n");
 
#9 lore raw
In reply to Peter Zijlstra
Quoted 8 lines · Peter Zijlstra
Quoted 4 lines · Jaswinder Singh Rajput

With fresh copy, perfstat is working.

But kerneltop is still not working and I am not getting any NMI Interrupts.

What does your dmegs look like?

Quoted 8 lines

Mine read like:

[    0.012002] AMD Performance Monitoring support detected.
[    0.013002] ... num counters:    4
[    0.014001] ... value mask:      0000ffffffffffff
[    0.015001] ... fixed counters:  0
[    0.016001] ... counter mask:    000000000000000f
[    0.010998] AMD Performance Monitoring support detected.
[    0.011006] ... num counters:    4
[    0.011301] ... value mask:      0000000000000086
[    0.011581] ... fixed counters:  0
[    0.011808] ... counter mask:    000000000000000f
[    0.011998] ACPI: Core revision 20081204
Quoted 4 lines

and its generating IRQs:

CNT: 168612 166040 Performance counter interrupts

NMI:          0          0   Non-maskable interrupts
LOC:        213     980394   Local timer interrupts
CNT:          0          0   Performance counter interrupts
RES:     125349     123542   Rescheduling interrupts

--
JSR

#11 lore raw
In reply to Peter Zijlstra
Quoted 8 lines · Peter Zijlstra
Quoted 1 line · Jaswinder Singh Rajput

But kerneltop is still not working and I am not getting any NMI Interrupts.

Does the below work for you?

The BKGD suggests its 48 for all existing machines (it certainly is for mine).

Yeah, this did the trick on AMD/Barcelona!

NMI: 11310 14932 11729 12705 11207

13454      13378      10714      14526      11986      10973      
10765      12669      12093      13430       9661   Non-maskable 

interrupts

KernelTop works just fine:

------------------------------------------------------------------------------
KernelTop: 1019 irqs/sec kernel:27.9% [NMI, 1000000 CPU
cycles], (all, 16 CPUs)
------------------------------------------------------------------------------

events RIP kernel function
______ ______ ________________ _______________

            5260.00 - ffffffff8037e420 : clear_page_c
            1065.00 - ffffffff805811a0 : page_fault
             699.00 - ffffffff8037e690 : copy_user_generic_string!
             669.00 - ffffffff80294ed8 : get_page_from_freelist
             575.00 - ffffffff804cd033 : acpi_pm_read
             551.00 - ffffffff8025fd1b : generic_exec_single
             416.00 - ffffffff80580c9c : _spin_lock
             413.00 - ffffffff802a32ad : unmap_vmas
             392.00 - ffffffff80580fc2 : _spin_lock_irqsave

Thanks!

Ingo

#12 lore raw
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Peter Zijlstra

On Fri, 2009-03-06 at 00:26 +0530, Jaswinder Singh Rajput wrote:

Quoted 9 lines
Quoted 6 lines
Quoted 1 line

But kerneltop is still not working and I am not getting any NMI Interrupts.

Does the below work for you?

The BKGD suggests its 48 for all existing machines (it certainly is for mine).

Yeah, this did the trick on AMD/Barcelona!

Nice.

but on my AMD box I got interrupts and NMI for some time:

NMI:        100    1108209   Non-maskable interrupts
LOC:        213     192815   Local timer interrupts
CNT:        100    1108206   Performance counter interrupts
RES:      21132      20397   Rescheduling interrupts

but kerneltop does not show anything, it only shows # irqs/sec for some time and then irqs also become 0.

Here is demg:

kerneltop

[  196.677297] Uhhuh. NMI received for unknown reason 00 on CPU 1.
[  196.677297] Do you have a strange power saving mode enabled?
[  196.677297] Dazed and confused, but trying to continue
[  198.854293] PERFMON: max interrupts exceeded!
[  198.865302] PERFMON: max interrupts exceeded!
[  198.866304] PERFMON: max interrupts exceeded!
[  206.066271] PERFMON: max interrupts exceeded!
[  206.067304] PERFMON: max interrupts exceeded!
[  206.068265] PERFMON: max interrupts exceeded!
[  211.176291] PERFMON: max interrupts exceeded!
[  211.177302] PERFMON: max interrupts exceeded!
[  211.235304] PERFMON: max interrupts exceeded!
[  211.236296] PERFMON: max interrupts exceeded!
[  214.245302] PERFMON: max interrupts exceeded!
[  225.257276] PERFMON: max interrupts exceeded!
[  225.258271] PERFMON: max interrupts exceeded!
[  297.385301] PERFMON: max interrupts exceeded!
[  297.386290] PERFMON: max interrupts exceeded!
[  310.405304] PERFMON: max interrupts exceeded!
[  392.677261] Uhhuh. NMI received for unknown reason 00 on CPU 1.
[  392.677261] Do you have a strange power saving mode enabled?
[  392.677261] Dazed and confused, but trying to continue
[  394.837275] PERFMON: max interrupts exceeded!
[  394.848268] PERFMON: max interrupts exceeded!
[  394.849266] PERFMON: max interrupts exceeded!
[  396.891309] PERFMON: max interrupts exceeded!
[  396.892294] PERFMON: max interrupts exceeded!
[  397.415302] PERFMON: max interrupts exceeded!
[  397.594299] PERFMON: max interrupts exceeded!
[  398.395302] PERFMON: max interrupts exceeded!
[  398.919304] PERFMON: max interrupts exceeded!
[  403.839309] __ratelimit: 1 callbacks suppressed
[  403.839319] PERFMON: max interrupts exceeded!
[  403.840302] PERFMON: max interrupts exceeded!
[  403.841298] PERFMON: max interrupts exceeded!
[  403.842297] PERFMON: max interrupts exceeded!
[  403.843267] PERFMON: max interrupts exceeded!
[  403.853303] PERFMON: max interrupts exceeded!
[  403.854290] PERFMON: max interrupts exceeded!
[  403.855301] PERFMON: max interrupts exceeded!
[  403.857301] PERFMON: max interrupts exceeded!
[  403.858304] PERFMON: max interrupts exceeded!
[  409.097310] __ratelimit: 13 callbacks suppressed
[  409.097321] PERFMON: max interrupts exceeded!
[  409.098299] PERFMON: max interrupts exceeded!
[  411.138303] PERFMON: max interrupts exceeded!
[  413.177301] PERFMON: max interrupts exceeded!
[  413.178289] PERFMON: max interrupts exceeded!
[  414.442311] PERFMON: max interrupts exceeded!
[  415.217304] PERFMON: max interrupts exceeded!
[  415.218291] PERFMON: max interrupts exceeded!
[  417.257307] PERFMON: max interrupts exceeded!
[  417.258294] PERFMON: max interrupts exceeded!
[  418.585295] PERFMON: max interrupts exceeded!
[  418.600293] PERFMON: max interrupts exceeded!
[  427.436301] PERFMON: max interrupts exceeded!
[  427.437307] PERFMON: max interrupts exceeded!
[  429.477269] PERFMON: max interrupts exceeded!
[  429.479264] PERFMON: max interrupts exceeded!
[  431.514269] PERFMON: max interrupts exceeded!
[  431.516294] PERFMON: max interrupts exceeded!
[  433.557303] PERFMON: max interrupts exceeded!
[  433.558290] PERFMON: max interrupts exceeded!
[  435.585305] PERFMON: max interrupts exceeded!
[  435.586291] PERFMON: max interrupts exceeded!
[  435.620276] PERFMON: max interrupts exceeded!
[  437.619294] PERFMON: max interrupts exceeded!
[  437.620295] PERFMON: max interrupts exceeded!
[  437.621289] PERFMON: max interrupts exceeded!
[  437.622290] PERFMON: max interrupts exceeded!
[  439.637302] PERFMON: max interrupts exceeded!
[  439.638291] PERFMON: max interrupts exceeded!
[  439.645301] PERFMON: max interrupts exceeded!
[  439.646298] PERFMON: max interrupts exceeded!
[  441.684302] PERFMON: max interrupts exceeded!
[  441.685298] PERFMON: max interrupts exceeded!
[  456.974308] PERFMON: max interrupts exceeded!
[  466.542303] PERFMON: max interrupts exceeded!
[  468.924294] PERFMON: max interrupts exceeded!
[  471.246282] PERFMON: max interrupts exceeded!
[  471.257269] PERFMON: max interrupts exceeded!
[  471.258268] PERFMON: max interrupts exceeded!
[  473.295297] PERFMON: max interrupts exceeded!
[  473.296301] PERFMON: max interrupts exceeded!
[  475.312292] PERFMON: max interrupts exceeded!
[  475.323302] PERFMON: max interrupts exceeded!
[  475.324302] PERFMON: max interrupts exceeded!
[  477.346303] PERFMON: max interrupts exceeded!
[  477.347300] PERFMON: max interrupts exceeded!
[  479.381303] PERFMON: max interrupts exceeded!
[  480.685269] PERFMON: max interrupts exceeded!
[  480.686270] PERFMON: max interrupts exceeded!
[  481.405307] PERFMON: max interrupts exceeded!
[  483.425304] PERFMON: max interrupts exceeded!
[  483.426291] PERFMON: max interrupts exceeded!
[  485.461306] PERFMON: max interrupts exceeded!
[  485.462293] PERFMON: max interrupts exceeded!
[  487.499304] PERFMON: max interrupts exceeded!
[  487.500290] PERFMON: max interrupts exceeded!
[  489.524279] PERFMON: max interrupts exceeded!
[  491.546304] PERFMON: max interrupts exceeded!
[  491.547292] PERFMON: max interrupts exceeded!
[  491.548285] PERFMON: max interrupts exceeded!
[  493.581307] PERFMON: max interrupts exceeded!
[  493.582301] PERFMON: max interrupts exceeded!
[  495.605303] PERFMON: max interrupts exceeded!
[  495.606311] PERFMON: max interrupts exceeded!
[  497.641303] PERFMON: max interrupts exceeded!
[  499.665303] PERFMON: max interrupts exceeded!
[  499.666325] PERFMON: max interrupts exceeded!
[  501.699295] PERFMON: max interrupts exceeded!
[  501.700301] PERFMON: max interrupts exceeded!
[  503.715304] PERFMON: max interrupts exceeded!
[  503.716292] PERFMON: max interrupts exceeded!
[  503.717309] PERFMON: max interrupts exceeded!
[  503.725302] PERFMON: max interrupts exceeded!
[  505.748271] PERFMON: max interrupts exceeded!
[  505.749382] PERFMON: max interrupts exceeded!
[  507.782304] PERFMON: max interrupts exceeded!
[  507.783291] PERFMON: max interrupts exceeded!
[  509.804303] PERFMON: max interrupts exceeded!
[  509.805302] PERFMON: max interrupts exceeded!
[  511.840303] PERFMON: max interrupts exceeded!
[  511.841299] PERFMON: max interrupts exceeded!
[  515.565261] Uhhuh. NMI received for unknown reason 00 on CPU 1.
[  515.565261] Do you have a strange power saving mode enabled?
[  515.565261] Dazed and confused, but trying to continue
[  515.569284] PERFMON: max interrupts exceeded!
[  515.570372] PERFMON: max interrupts exceeded!
[  515.571355] PERFMON: max interrupts exceeded!
[  515.574304] PERFMON: max interrupts exceeded!
[  515.575287] PERFMON: max interrupts exceeded!
[  515.576283] PERFMON: max interrupts exceeded!
[  515.577289] PERFMON: max interrupts exceeded!
[  515.579318] PERFMON: max interrupts exceeded!
[  516.948088] __ratelimit: 113 callbacks suppressed
[  516.948118] PERFMON: max interrupts exceeded!
[  516.951315] PERFMON: max interrupts exceeded!
[  516.952298] PERFMON: max interrupts exceeded!
[  516.953315] PERFMON: max interrupts exceeded!
[  516.954406] PERFMON: max interrupts exceeded!
[  516.955390] PERFMON: max interrupts exceeded!
[  516.956336] PERFMON: max interrupts exceeded!
[  516.957277] PERFMON: max interrupts exceeded!
[  516.958285] PERFMON: max interrupts exceeded!
[  516.959313] PERFMON: max interrupts exceeded!
[  522.039265] __ratelimit: 248 callbacks suppressed
[  522.039294] PERFMON: max interrupts exceeded!
[  522.040321] PERFMON: max interrupts exceeded!
[  522.042299] PERFMON: max interrupts exceeded!
[  522.043287] PERFMON: max interrupts exceeded!
[  522.044277] PERFMON: max interrupts exceeded!
[  522.045299] PERFMON: max interrupts exceeded!
[  522.048017] PERFMON: max interrupts exceeded!
[  522.148095] PERFMON: max interrupts exceeded!
[  522.149279] PERFMON: max interrupts exceeded!
[  522.151350] PERFMON: max interrupts exceeded!
[  527.043304] __ratelimit: 1024 callbacks suppressed
[  527.043314] PERFMON: max interrupts exceeded!
[  527.044303] PERFMON: max interrupts exceeded!
[  527.045298] PERFMON: max interrupts exceeded!
[  527.046291] PERFMON: max interrupts exceeded!
[  527.047271] PERFMON: max interrupts exceeded!
[  527.048276] PERFMON: max interrupts exceeded!
[  527.049303] PERFMON: max interrupts exceeded!
[  527.052303] PERFMON: max interrupts exceeded!
[  527.053566] PERFMON: max interrupts exceeded!
[  527.054304] PERFMON: max interrupts exceeded!
[  532.049312] __ratelimit: 2191 callbacks suppressed
[  532.049321] PERFMON: max interrupts exceeded!
[  532.050303] PERFMON: max interrupts exceeded!
[  532.052298] PERFMON: max interrupts exceeded!
[  532.053305] PERFMON: max interrupts exceeded!
[  532.065291] PERFMON: max interrupts exceeded!
[  532.067263] PERFMON: max interrupts exceeded!
[  532.069299] PERFMON: max interrupts exceeded!
[  532.071300] PERFMON: max interrupts exceeded!
[  532.073311] PERFMON: max interrupts exceeded!
[  532.075302] PERFMON: max interrupts exceeded!
[  537.062306] __ratelimit: 1414 callbacks suppressed
[  537.062335] PERFMON: max interrupts exceeded!
[  537.063443] PERFMON: max interrupts exceeded!
[  537.066285] PERFMON: max interrupts exceeded!
[  537.067289] PERFMON: max interrupts exceeded!
[  537.068282] PERFMON: max interrupts exceeded!
[  537.069322] PERFMON: max interrupts exceeded!
[  537.071312] PERFMON: max interrupts exceeded!
[  537.072295] PERFMON: max interrupts exceeded!
[  537.073279] PERFMON: max interrupts exceeded!
[  537.074287] PERFMON: max interrupts exceeded!

[root@hpdv5 new]#

--
JSR

#13 lore raw
Quoted 19 lines · Jaswinder Singh Rajput
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Peter Zijlstra

On Fri, 2009-03-06 at 00:26 +0530, Jaswinder Singh Rajput wrote:

Quoted 9 lines
Quoted 6 lines
Quoted 1 line

But kerneltop is still not working and I am not getting any NMI Interrupts.

Does the below work for you?

The BKGD suggests its 48 for all existing machines (it certainly is for mine).

Yeah, this did the trick on AMD/Barcelona!

Nice.

but on my AMD box I got interrupts and NMI for some time:

Quoted 7 lines
NMI:        100    1108209   Non-maskable interrupts
LOC:        213     192815   Local timer interrupts
CNT:        100    1108206   Performance counter interrupts
RES:      21132      20397   Rescheduling interrupts

but kerneltop does not show anything, it only shows # irqs/sec for some time and then irqs also become 0.

Hm. Kerneltop on my Q6600 has ceased to function with n 1 as well, but n 0 still works. Zero NMIs.

Perfstat has also gone south.
perfstat: perfstat.c:423: main: Assertion `res == sizeof(single_count)' failed.

-Mike

#14 lore raw
Quoted 25 lines · Mike Galbraith
Quoted 24 lines · Jaswinder Singh Rajput
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Peter Zijlstra

On Fri, 2009-03-06 at 00:26 +0530, Jaswinder Singh Rajput wrote:

Quoted 9 lines
Quoted 6 lines
Quoted 1 line

But kerneltop is still not working and I am not getting any NMI Interrupts.

Does the below work for you?

The BKGD suggests its 48 for all existing machines (it certainly is for mine).

Yeah, this did the trick on AMD/Barcelona!

Nice.

but on my AMD box I got interrupts and NMI for some time:

NMI:        100    1108209   Non-maskable interrupts
LOC:        213     192815   Local timer interrupts
CNT:        100    1108206   Performance counter interrupts
RES:      21132      20397   Rescheduling interrupts
Quoted 8 lines
Quoted 2 lines

but kerneltop does not show anything, it only shows # irqs/sec for some time and then irqs also become 0.

Hm. Kerneltop on my Q6600 has ceased to function with n 1 as well, but n 0 still works. Zero NMIs.

Perfstat has also gone south.
perfstat: perfstat.c:423: main: Assertion `res == sizeof(single_count)' failed.

Ah, my woes are gonna be 2743a5b0fa6f309da904f2190a9cc25deee34dbd.

Never mind.

-Mike

#15 lore raw
Quoted 1 line · Mike Galbraith

Ah, my woes are gonna be 2743a5b0fa6f309da904f2190a9cc25deee34dbd.

(and indeed, they both work fine now)

#3 lore raw
In reply to Peter Zijlstra
Quoted 31 lines · Peter Zijlstra
Commit-ID:  b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Gitweb:     http://git.kernel.org/tip/b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Author:     "Peter Zijlstra" <[email protected]>
AuthorDate: Thu, 5 Mar 2009 18:08:27 +0100
Commit:     Ingo Molnar <[email protected]>
CommitDate: Thu, 5 Mar 2009 18:25:16 +0100

perfcounters: IRQ and NMI support on AMD CPUs

The below completes the K7+ performance counter support:

- IRQ support
- NMI support

KernelTop output works now as well.

Cc Jaswinder Singh Rajput <[email protected]>
Cc Paul Mackerras <[email protected]>

LKML-Reference: <1236273633.5187.286.camel@laptop>

---
 arch/x86/kernel/cpu/perf_counter.c |  272 ++++++++++++++++++++++++++++++------
 1 files changed, 228 insertions(+), 44 deletions(-)
arch/x86/kernel/cpu/perf_counter.c +0 -0
index 3b65f19..6ebe9ab 100644

[...]

arch/x86/kernel/cpu/perf_counter.c static u64 pmc_intel_save_disable_all(void) · line 251

Quoted 12 lines
arch/x86/kernel/cpu/perf_counter.c +6 -2
@@ -205,18 +243,24 @@ static u64 pmc_intel_save_disable_all(void)
 
 static u64 pmc_amd_save_disable_all(void)
 {
-	int idx;
-	u64 val, ctrl = 0;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	int enabled, idx;
+
+	enabled = cpuc->enabled;
+	cpuc->enabled = 0;
+	barrier();

Peter,

please add comments to all barrier()s you added to this file. Is it sufficient in hw_perf_save_disable() to stop NMI's only on the current core?

-Robert

Quoted 20 lines
arch/x86/kernel/cpu/perf_counter.c +7 -5
 	for (idx = 0; idx < nr_counters_generic; idx++) {
+		u64 val;
+
 		rdmsrl(MSR_K7_EVNTSEL0 + idx, val);
-		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE)
-			ctrl |= (1 << idx);
-		val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
-		wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		if (val & ARCH_PERFMON_EVENTSEL0_ENABLE) {
+			val &= ~ARCH_PERFMON_EVENTSEL0_ENABLE;
+			wrmsrl(MSR_K7_EVNTSEL0 + idx, val);
+		}
 	}
 
-	return ctrl;
+	return enabled;
 }
 
 u64 hw_perf_save_disable(void)

[...]

Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]
#4 lore raw
Quoted 50 lines · Robert Richter
Quoted 31 lines · Peter Zijlstra
Commit-ID:  b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Gitweb:     http://git.kernel.org/tip/b0f3f28e0f14eb335f67bfaae33ce8b8d74fd58b
Author:     "Peter Zijlstra" <[email protected]>
AuthorDate: Thu, 5 Mar 2009 18:08:27 +0100
Commit:     Ingo Molnar <[email protected]>
CommitDate: Thu, 5 Mar 2009 18:25:16 +0100

perfcounters: IRQ and NMI support on AMD CPUs

The below completes the K7+ performance counter support:

- IRQ support
- NMI support

KernelTop output works now as well.

Cc Jaswinder Singh Rajput <[email protected]>
Cc Paul Mackerras <[email protected]>

LKML-Reference: <1236273633.5187.286.camel@laptop>

---
 arch/x86/kernel/cpu/perf_counter.c |  272 ++++++++++++++++++++++++++++++------
 1 files changed, 228 insertions(+), 44 deletions(-)
arch/x86/kernel/cpu/perf_counter.c +0 -0
index 3b65f19..6ebe9ab 100644

[...]

arch/x86/kernel/cpu/perf_counter.c static u64 pmc_intel_save_disable_all(void) · line 251

Quoted 12 lines
arch/x86/kernel/cpu/perf_counter.c +6 -2
@@ -205,18 +243,24 @@ static u64 pmc_intel_save_disable_all(void)
 
 static u64 pmc_amd_save_disable_all(void)
 {
-	int idx;
-	u64 val, ctrl = 0;
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	int enabled, idx;
+
+	enabled = cpuc->enabled;
+	cpuc->enabled = 0;
+	barrier();

Peter,

Quoted 3 lines

please add comments to all barrier()s you added to this file. Is it sufficient in hw_perf_save_disable() to stop NMI's only on the current core?

Yes, counters are fundamentally per cpu. But you're right, I'll go through the code and add comments and fix up some barriers.